After more than a decade of explosive growth and eight previous re:Invent conferences, cloud watchers are used to the annual firehose of information packed into CEO Andy Jassy’s keynotes.
However, the zeitgeist of 2020 thwarted the throngs that would normally pack into the Sands Convention Center and relegated Jassy to an empty facsimile of the typical expo hall and stage.
Regardless, Jassy didn’t disappoint, regaling a larger-than-usual online audience with dozens of announcements spread over a three-hour tour de force of vocal, adrenal and vesical stamina.
With revenue of almost $130 million per day and still doubling every two-three years, AWS has grown to resemble its retail parent: the cloud service with something for everyone (for more earnings context, see: AWS Q3 2020 earnings release, and slides - PDF link). As re:Invent has expanded, it becomes progressively more difficult to identify overarching themes in the dozens of product announcements and updates. Indeed, finding significant patterns amidst the barrage of keynote slides and AWS blog posts has become a Rorschach test for cloud watchers: what you highlight is more a reflection of personal preferences and biases than AWS priorities. Nonetheless, I’ll highlight some topics I found significant, rationalizing a few by noting their prominent position early in Jassy’s address, before audience fatigue and temporal distractions took over.
Building modern apps the AWS way
I expected AWS to emphasize two areas, homegrown hardware and serverless services, indeed got top billing in Jassy’s keynote. With serverless — which, for the clueless and cynics out doesn’t mean devoid of computational hardware, but instead services that are used and decommissioned on-demand automatically and without previously provisioning instances, storage or capacity - AWS is emulating Google by expanding the definition beyond event-driven functions like Lambda to include managed applications and platforms, including an API gateway, pub/sub notifications and message queuing, an event bus and workflow scheduler. This year, Jassy emphasized its Aurora DBMS service, where a new version brings faster performance and scaling along with better support for SQL Server applications.
Jassy claims that Aurora, its Oracle killer that is now being aimed at Microsoft SQL Server customers, is the fastest-growing service in its history. With version 2, Aurora can scale to “hundreds of thousands of transactions in a fraction of a second” without requiring customers to pre-provision peak capacity. AWS estimates that Aurora’s improved-granularity auto-scaling cuts cost up to 90% compared with traditional cloud databases. AWS also introduced Babelfish for Aurora PostgreSQL, a code translation layer that can parse SQL Server’s network protocol and command, making it interoperable with existing SQL Server drivers and applications. AWS also open sourced the Bablefish code to facilitate the migration of SQL Server developers and tools to Aurora.
Containers and serverless functions are the foundation of modern applications and AWS isn’t about to let Google get a significant lead in mindshare or technology. Containers are the vehicle for custom applications and services with Kubernetes now the preferred workload and cluster management platform. AWS has long hedged its bets by offering two managed orchestration systems, ECS (native, using EC2 instances) and EKS (Kubernetes) and announced forthcoming support for running them in a hybrid configuration on internal bare metal or VMware servers. Both remain managed services (CaaS) that use the same configuration and management UI as cloud-based alternatives.
More interesting is how AWS continues to expand Lambda into a full-fledged execution environment. A problem with trying to make sophisticated Lambda functions are the required system dependencies like libraries and runtime environments. These can now be accommodated by Lambda’s support for container images as large as 10GB. AWS has built base images with Python, Node.js, Java, .NET, Go and Ruby, but supports custom runtimes by bundling the requisite components and the Lambda Runtime API into an Amazon Linux image.
AWS also made Lambda more attractive for frequently-used, short-duration workloads by reducing the billing granularity from 100 to 1ms. AWS provided an example of a function in a 100,000-user web app that is called 20-times per day per client. Although the hypothetical code only takes 28ms to execute, the old pricing model rounded it up to 100ms, costing almost 3-times as much per month as the new billing scheme. Together, these improvements make Lambda significantly more compelling than a persistent VM for glue logic, scheduled jobs or high-use, short-duration functions requiring low-latency and high scalability. Jassy noted that almost half of the new applications deployed on AWS this year use Lambda and collectively run over a million transactions per second.
Taking a page from Apple’s playbook - AWS likes silicon too
This year marks a turning point in the adoption of specialized, often custom-designed processors and SoCs in favor of general-purpose CPUs. Apple continues to lead the way in customized consumer components, with its A14 smoking competitive SoCs from Qualcomm and Samsung and new M1 chip outperforming “everything Intel has to offer” on the PC side. AWS followed suit as the first to announce a custom Arm SoC for cloud workloads with its Graviton instances two years ago. Last year it made significant improvements with the Graviton 2, which got a speed boost this year.
The C6g instances target compute-heavy workloads by improving network bandwidth 4-times to 100 Gbps and block storage (EBS) throughput to 38Gbps. Overall, Jassy claims that the Graviton2 family of instances deliver “40% better price-performance for all workloads” than Xeon-based alternatives. He also highlighted the broad Graviton compatibility across the AWS service portfolio, including ECS, EKS and CodeX developer services, and support by third-party vendors and Linux distros. Indeed, he noted that customers are often surprised by how quickly they can port applications to Graviton.
AWS introduced its first custom chip for machine learning (ML) inference calculations two years ago with Inferentia, leaving NVIDIA V100 GPUs in its P3 instances as the preferred option for model training, until now. Jassy announced two new offerings, the homegrown AWS Trainium and one using Intel’s Habana Labs Gaudi processor, designed to provide better price performance. AWS didn’t offer technical or performance details about the Trainium, which won’t be available until sometime next year, but says it is compatible with the Neuron SDK used to develop Inferentia models. Since Neuron includes interfaces for popular ML frameworks like TensorFlow and MXNet, AWS expects developers to have little difficulty moving model training from GPU to Trainium instances.
Habana’s Gaudi, which Intel acquired in 2019, is another special-purpose processor designed for AI model training that uses an approach similar to Google’s TPU with eight tensor (vector) processing cores per chip. AWS will bundle 8 Gaudi accelerators in each instance, which it expects to “deliver up to 40% better price performance than current GPU-based EC2 instances for training deep learning models.” Like Neuron, the Gaudi SDK supports most popular AI frameworks. Both Trainium and Gaudi instances will be usable in EKC and ECS clusters or with the SageMaker development platform.
In reviewing another successful year, Jassy touted AWS as the broadest and deepest cloud platform and one glance at the accompanying eye chart shows that he isn’t exaggerating. Although AWS does have a service for everyone, data from the cloud consultancy 2ndWatch shows that core IaaS products like EC2, RDS, DynamoDB and S3 remain the most popular services. However, 2ndWatch also finds that newer SaaS and platform services like Transcribe, Comprehend, Personalize and Athena have the fastest uptake. While these services often get lost amidst the annual flood of re:Invent announcements, AWS has found and filled a latent need for packaged, automated platforms and applications that relieve IT and developers from the burden of provisioning and administering cloud services.
It was telling that although Jassy led his keynote with the meat and potatoes of compute, containers, storage and components, he spent the latter half discussing SaaS business products like AWS Connect (contact center), QuickSight (BI), Glue (ETL), Lookout (image detection), Monitron (predictive analytics) and DevOps Guru (code analysis). I’ll have more to say about these later, but collectively they’re turning AWS into a cloud supermart that defies traditional classifications.
My focus here is the increasingly differentiated foundation AWS has built for its core infrastructure services, which are creating the same sort of competitive moat that parent Amazon’s has built through its sustained investment in distribution and logistics infrastructure. AWS has recreated the virtuous cycle Bezos first envisioned for Amazon and the resulting flywheel effect poses nearly insurmountable challenges for AWS’s cloud competitors.